Historical Perspective: Why PIM Now and What Happened in the Past

Kevin Skadron
University of Virginia

Abstract

Applications are increasingly data-intensive and bound by the performance of the memory and/or storage system. This “memory wall” arises from several factors: the volume of data is increasing exponentially, outstripping cache capacities; many applications extensively use streaming data with little or no temporal reuse; as algorithms become more sophisticated, access patterns are often unfriendly to effective caching; and the computation intensity of many of these algorithms is low–we often spend more time and energy moving data to the processor than we spend computing on the data.  All these factors motivate breaking down the classic von Neumann architecture that separates processing and memory, and computing as close to the data as possible, with processing elements either tightly coupled with memory or storage, or possibly even embedded directly in the memory chips, i.e. processing in memory or PIM. 

 The memory wall has been a concern for decades, with numerous proposals over the years for processing-in-memory and near-data architectures.  This talk will review the motivation for processing in memory and some prior proposals, then provide an overview of the current landscape, take a deep dive into the DRAM-CAM in-memory pattern-matching accelerator developed by my research group, and then conclude with some observations about possible directions for PIM.

[Slides]

Bio

Kevin Skadron is the Harry Douglas Forsyth Professor of Computer Science at the University of Virginia, where he has been on the faculty since 1999, after receiving his PhD at Princeton. He served as department chair from 2012-2021. He is also director of the Center for Research on Intelligent Storage and Processing in Memory, part of the SRC JUMP program, as well as director of the Center for Automata Processing. He is a Fellow of the IEEE and the ACM, a recipient of the 2011 ACM SIGARCH Maurice Wilkes Award, and co-founder of IEEE Computer Architecture Letters. Skadron‘s research interests include design and application of accelerators and heterogeneous architectures, their memory hierarchies, and associated power, thermal, reliability, and programming challenges. He and his colleagues and students have developed a number of tools to support research on these topics, such as MNCaRT (for automata processing), HotSpot, and Rodinia.